A New Study Shows That Chatgpt Can Predict Stock Moves Based On Headlines

The first wave of academic research bringing ChatGPT to the world of finance is on its way, and based on preliminary findings, the recent hoopla is warranted.

‍

This month, two additional studies were released that used the artificial intelligence chatbot to perform market-relevant tasks: one in assessing whether Federal Reserve remarks were hawkish or dovish, and the other in determining whether headlines were good or bad for a company.

‍

ChatGPT passed both tests, indicating a potentially significant advancement in the use of technology to convert reams of text ranging from news articles to tweets and speeches into trading signals.

‍

Of course, that process is nothing new on Wall Street, where quants have long utilized the type of language models powering the chatbot to guide a variety of tactics. Yet, the findings indicate that OpenAI's technology has advanced to a new level in terms of processing subtlety and context.

‍

"It's one of the few occasions where the hype is true," said Slavi Marinov, head of machine learning at Man AHL, which has been employing natural language processing (NLP) technology for years to interpret documents such as earnings transcripts and Reddit postings.

‍

The first paper, titled "Can ChatGPT Decipher Fedspeak?" ", two Fed researchers discovered that ChatGPT was the closest to humans in determining whether the central bank's utterances were dovish or hawkish. The Richmond Fed's Anne Lundgaard Hansen and Sophia Kazinnik demonstrated that it outperformed a regularly used Google model called BERT as well as classifications based on dictionaries.

‍

ChatGPT was even able to explain its classifications of Fed policy pronouncements in a way that mirrored the central bank's analyst, who also read the language to provide a human baseline for the research.

‍

Similar to an intelligent analyst

The robot claimed the phrase was dovish because it implied the economy had not fully recovered yet. For example, take this statement from a May 2013 statement: "Labor market conditions have seen some improvement in recent months, on balance, although the unemployment rate remains elevated." That was the analyst's conclusion as well; Bryson, a 24-year-old man "recognized for his brilliance and curiosity," was mentioned in the study.

‍

Can ChatGPT Predict Stock Price Movements? is the second research. ChatGPT was instructed by Alejandro Lopez-Lira and Yuehua Tang at the University of Florida to pretend to be a financial expert and read business news headlines in their paper, "Return Predictability with Big Language Models". They utilized news from a time frame beyond late 2021, which the chatbot's training data didn't cover.

‍

The research discovered a statistical relationship between the responses provided by ChatGPT and the stock's subsequent movements, indicating that the technology was able to accurately comprehend the significance of the news.

In response to the question of whether Oracle would benefit or suffer from the headline "Rimini Street Fined $630,000 in Case Against Oracle," ChatGPT stated that the latter was true because the sanction "could potentially boost investor confidence in Oracle's ability to protect its intellectual property and increase demand for its products and services."

‍

Using NLP to determine a stock's popularity from Twitter or to include the most recent news about a firm is now almost routine for the most proficient quants. Yet, the innovations shown by ChatGPT appear to be opening up whole new information and making the technology more approachable to a larger group of financial professionals.

‍

Speeding up the whole process

According to Mr. Marinov, while it's no surprise that robots can already read almost as well as humans, ChatGPT has the ability to accelerate the entire process.

‍

When Man AHL was originally developing the models, the quant hedge fund manually labeled each statement as positive or negative for an asset to provide a blueprint for the robots to read the language. The London-based business then converted the entire process into a game, ranking participants and calculating how much they agreed on each statement, allowing all employees to participate.

‍

According to the two recent articles, ChatGPT can perform similar tasks without being properly taught. According to Fed research, this so-called zero-shot learning already outperforms previous technology, but fine-tuning it based on some specific cases made it much better.

‍

"Previously, you had to categorize the data manually," Mr. Marinov, who previously co-founded an NLP startup, explained. "You might now supplement that by developing the appropriate ChatGPT prompt."

‍